AI News5 min read

Copilot in October: assisted permissions, model policy API and two retirement dates

Direct answer

Between October 15 and October 26, 2026 GitHub made assisted permissions generally available in Copilot CLI, the Copilot app and VS Code, published model policy endpoints for enterprises, teams and organizations, and began enforcing the global default policy. A separate changelog entry schedules Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra for retirement on November 16.

GitHub published four administrative changes to Copilot in the two weeks before October 26, 2026. Three of them add control points; the fourth is a deadline. None of them changes what a model can do, but together they change what an administrator can prove after an incident.

We read the four changelog entries on October 28, 2026. We did not deploy these settings in an enterprise tenant and did not take a model through its retirement date.

Assisted permissions are now generally available

The October 26 entry makes assisted permissions generally available in Copilot CLI, the Copilot app and Visual Studio Code. The mechanism: when Copilot requests permission to run a command, access a file system path, or open a URL, a large language model evaluates the request. Requests classified as clearly safe proceed automatically; the rest require approval.

It is enabled per surface, not globally:

Surface Where to switch it on
Copilot CLI Run /permissions, then select assisted
Copilot app Open the mode picker, click Tool permissions, then Assisted
VS Code Open the permissions picker, then Assisted permissions

The important part for a regulated team is the administrative layer. Enterprise-managed settings can block Allow All while keeping Assisted available, or disable Assisted separately. Administrators can also choose the default permission mode for new sessions, and developers can still select another mode that policy permits. If the default is unavailable or prohibited, the session starts with manual approval.

Read that last sentence slowly, because it is the failure mode worth testing. A policy that removes the default does not silently downgrade to Allow All; it falls back to manual approval. Our managed permissions explainer covers the deny, ask and allow precedence that sits underneath this, and the September update it documents still applies.

The model policy API covers four levels

The October 19 entry makes GET and PATCH endpoints generally available for model policies at four levels: enterprise, enterprise team, agent app, and organization. Enterprise policies can delegate model control to organizations in organization mode, or to teams and apps in enterprise-teams mode.

For anyone maintaining a policy as code, this is the change that removes a manual step. Note what the delegation means in practice: the effective policy for a developer is the result of an inheritance chain, not a single setting. Before you script a rollout, decide which level owns each decision, then confirm that a lower level cannot widen what a higher level restricted.

The global default policy began enforcing

The October 22 entry says GitHub is gradually rolling out enforcement of the global default policy through October 26, so different enterprises and organizations may see it take effect at different times. Three details matter:

  • Eligible generally available features left unconfigured are set according to the selected global default of Enabled, Disabled, or Let organizations decide.
  • It is a one-time change. Later changes to the default policy will not move feature policies that are already set.
  • Explicit choices are preserved, and preview features remain opt-in.

That combination produces an audit question worth answering now: after the rollout, which features are set explicitly and which are merely inheriting a default? A feature enabled by inheritance looks identical in the UI to one enabled on purpose, and only one of them survives the next default change.

Two model retirement dates, one of them confirmed

The October 15 entry schedules retirements across all Copilot experiences, including Copilot Chat, inline edits, ask and agent modes, and code completions, for November 16, 2026:

Model Retirement date Suggested alternative
Claude Opus 4.8 2026-11-16 Claude Opus 5.5
Claude Opus 5 2026-11-16 Claude Opus 5.5
GPT-5.6 Terra 2026-11-16 GPT-6.1 Sol or GPT-6 Luna

One footnote limits the scope: Claude Opus 4.8 stays available to individual Copilot subscribers on annual plans, so they keep an Opus option. The others do not carry that exception.

Under default model enablement, the alternatives are enabled automatically for Copilot Enterprise and Business unless an administrator turned off the global default or explicitly disabled the model. If you turned the global default off, you have to enable the alternatives through their own model policies. No action is needed to remove the retired models.

The changelog also lists a separate retirement dated October 19, 2026. We did not read that entry’s model table, so this article makes no claim about which models it covers. Check it directly if your account still exposes older models.

The usage report now names the client

Two entries from October 22 and 23 change attribution rather than control. The usage metrics entry adds a totals_by_client array to the enterprise, organization, enterprise-user and organization-user 1-day and 28-day reports; each entry names the client, such as opencode, claude_code, codex, zed, copilot_cli or vscode, with request, session and token counts plus a model_selection_mode of user_selected, auto or product_selected. The October 22 entry replaces other and custom with the configured names of skills, custom agents, MCP servers, slash commands and plugins, and raises each list from five entries to 25.

model_selection_mode is the field to look at first. A request that the product chose is not evidence that a developer chose that model, and a cost discussion that mixes the two will assign the wrong decision to the wrong person.

What should an administrator check this week?

  1. List the models your teams actually use, then compare that list against the November 16 retirement table.
  2. Decide whether each affected workflow moves to the suggested alternative or to something you chose deliberately, and record the reason.
  3. Confirm which features are explicitly configured and which inherit the global default.
  4. Decide whether assisted permissions is on by default for new sessions, and whether Allow All stays available.
  5. Re-read the client support table before promising either permission mode on every surface.

None of these steps requires buying anything. All of them require an owner. The agent governance guide assigns those owners, and the pilot scorecard turns the answers into a testable checklist.

Frequently asked questions

What do Copilot assisted permissions decide?

When Copilot requests a command, a file path or a URL, a language model classifies the request. Requests judged clearly safe proceed; the rest still need your approval. GitHub made this generally available in Copilot CLI, the Copilot app and VS Code on October 26, 2026.

Can an administrator turn assisted permissions off?

GitHub documents enterprise-managed settings that can block Allow All while keeping Assisted available, or disable Assisted separately. Administrators can also set the default permission mode for new sessions, and developers keep whichever other mode policy permits.

Which Copilot models retire on November 16, 2026?

The changelog lists Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra, with Claude Opus 5.5 and GPT-6.1 Sol or GPT-6 Luna as suggested alternatives. Claude Opus 4.8 stays available to individual subscribers on annual plans.

Cookie settings

We use cookies only for analytics (GA4 + Matomo) to improve the docs. No ads, no tracking across sites.